EDNeRF: Editable Dynamic Neural Radiance Field Based on CUDA GPU Acceleration and Rendering Algorithms Improvements

NeRF (Neural Radiance Field) has emerged as a prominent research focus in recent years, demonstrating rapid advancement in the realms of 3D reconstruction and rendering. While the innovative NeRF or NeRF-type findings hold significant potential for practical application and scholarly value, it was important to note that initial iterations of NeRF are static. As such, dynamic and editable implementations of 3D reconstruction and rendering remained predominant in practical applications. Therefore, the pursuit of dynamic or editable NeRF or NeRF-like research held greater significance and represents the primary objective. In this work, an exceptionally detailed account of Dynamic NeRF was provided in the review, including its fundamental principles and derived advancements of the Gaussian Splatting and the Stable Video 3D (SV3D) that was based on Stable Diffusion model. Compared with other dynamic NeRF, the re-editable dynamic idea proposed in this thesis has more advanced practical performance and characterization. The main goal of thesis is achieving a method and the corresponding idea that can be used to gain a better performance in building a re-editable and dynamic NeRF-type model, and realizing the persistence of 3D generation and rendering operations. The literature review of the thesis covered detailed content and description about State-of-the-art NeRF models. In the research pipeline, there are three novel methods proposed to address the Research Objectives. First is to modify and enhance the structural design of the conventional Multi-layer perceptron (MLP) model to improve the 3D reconstruction and rendering performance. Next, the research proposed a series of CUDA GPU acceleration algorithms to improve the learning and processing speed. Finally, a novel editable dynamic NeRF model with higher realization complexity and robustness is introduced. Besides, compared with other dynamic NeRF-model, the proposed model realized the simple text-to-motions function and the persistence of model generation with innovative idea. Corresponding to the elaboration of the methodology, there were detailed experimental verification to prove the effectiveness of the proposed methods. A comprehensive analysis and outlook for the future works were also presented.<p></p>